def write_tfrecord(decode_path, write_path):
img_data = basis_code.basis_decode(decode_path)
#维度为6x10x3
img_data = tf.image.resize_images(img_data, [6, 10], method=1)
with tf.Session() as sess:
img_value = sess.run(img_data)
#总共写入多少文件
num_shards = 2
#每个文件写入多少数据
instance_per_shard = 3
for i in range(2):
#将文件分为多个文件。以便读取时方便匹配
filename = os.path.join(
write_path,
"data.tfrecords-" + str(i) + ".5d-of-" + str(num_shards) + ".5d")
writer = tf.python_io.TFRecordWriter(filename)
for j in range(instance_per_shard):
if i == 1:
j = j + 3
Log_Util.getlogger("write img" + str(j)).info(img_value[j])
img_raw = img_value[j].tostring()
example = tf.train.Example(features=tf.train.Features(
feature={
"labels": _int64_feature(j),
"img_data": _byte_feature(img_raw)
}))
try:
writer.write(example.SerializeToString())
except IOError:
print("write error!")
writer.close()
writer.close()
def adjust_brightness(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#将图像的亮度-0.5
adjust_less = tf.image.adjust_brightness(img_data, -0.5)
#色彩调整的API可能导致像素的实数值超出0.0-1.0的范围，因此在输出最终图像前需要将其值截断到0.0-1.0范围之间，否则不仅图像无法正常可视化，以此为输入的神经网络的训练质量也可能受到影响。
#如果对图像进行多项处理操作，那么这一截断过程应当在所有处理完成后进行。
adjust_less = tf.clip_by_value(adjust_less, 0.0, 1.0)
adjust_add = tf.image.adjust_brightness(img_data, 0.5)
adjust_add = tf.clip_by_value(adjust_add, 0.0, 1.0)
#在[-0.7, 0.7)的范围内随机调整图像的亮度
random_adjust = tf.image.random_brightness(img_data, 0.7)
random_adjust = tf.clip_by_value(random_adjust, 0.0, 1.0)
with tf.Session() as sess:
basis.drawing(sess.run(adjust_less))
basis.drawing(sess.run(adjust_add))
basis.drawing(sess.run(random_adjust))
basis_code.basis_encode(adjust_less,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(adjust_add,
basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(random_adjust,
basis_code.get_encode_path(filepath, 3))
def adjust_hue(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#下列依次将图像的色相加0.1、0.3、0.6、0.9
adjust_add1 = tf.image.adjust_hue(img_data, 0.1)
adjust_add2 = tf.image.adjust_hue(img_data, 0.3)
adjust_add3 = tf.image.adjust_hue(img_data, 0.6)
adjust_add4 = tf.image.adjust_hue(img_data, 0.9)
#在[0.3, 1)的范围内随机调整图像的色相
random_adjust = tf.image.random_hue(img_data, 0.3, 1)
with tf.Session() as sess:
basis.drawing(sess.run(adjust_add1))
basis.drawing(sess.run(adjust_add2))
basis.drawing(sess.run(adjust_add3))
basis.drawing(sess.run(adjust_add4))
basis.drawing(sess.run(random_adjust))
basis_code.basis_encode(adjust_add1,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(adjust_add2,
basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(adjust_add3,
basis_code.get_encode_path(filepath, 3))
basis_code.basis_encode(adjust_add4,
basis_code.get_encode_path(filepath, 4))
basis_code.basis_encode(random_adjust,
basis_code.get_encode_path(filepath, 5))
def adjust_standardization(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#将图像的亮度均值变为0，方差变为1
adjust = tf.image.per_image_standardization(img_data)
with tf.Session() as sess:
basis.drawing(sess.run(adjust))
basis_code.basis_encode(adjust, basis_code.get_encode_path(filepath, 1))
def method_two(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#以50%概率上下翻转图像
random_up_dowm_data = tf.image.random_flip_up_down(img_data)
# 以50%概率左右翻转图像
random_left_rigth = tf.image.random_flip_left_right(img_data)
with tf.Session() as sess:
basis.drawing(sess.run(random_up_dowm_data))
basis.drawing(sess.run(random_left_rigth))
basis_code.basis_encode(random_up_dowm_data,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(random_left_rigth,
basis_code.get_encode_path(filepath, 2))
def add_box(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#将图像缩小一点，这样可视化能让标注框更加清楚
img_data = tf.image.resize_images(img_data, [180, 267], method=1)
#tf.image.draw_bounding_boxes函数要求图像矩阵中的数字为实数，上面已经转换过。而且tf.image.draw_bounding_boxes函数的输入是一个batch的数据。
#也就是多张图像组成的四维矩阵，所以需要将解码之后的图像矩阵加一维
img_data = tf.expand_dims(img_data, 0)
#下面定义表示有两个标注框。一个标注框有4个数字，分别代表[Ymin, Xmin, Ymax, Xmax]。这里的数字都是图像的相对位置，比如在180x267的图像中。
#[0.33, 0.43, 0.48, 0.67]代表的大小为[0.33*180, 0.43*267, 0.48*180, 0.67*237]
boxes = tf.constant([[[0.13, 0.24, 0.55, 0.89], [0.33, 0.43, 0.48, 0.67]]])
result_data = tf.image.draw_bounding_boxes(img_data, boxes)
with tf.Session() as sess:
basis.drawing(sess.run(result_data[0]))
basis_code.basis_encode(result_data[0], basis_code.get_encode_path(filepath, 1))
def method_one(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#上下翻转
up_down_data = tf.image.flip_up_down(img_data)
#左右翻转
left_right_data = tf.image.flip_left_right(img_data)
#沿对角线翻转
transposed = tf.image.transpose_image(img_data)
with tf.Session() as sess:
basis.drawing(sess.run(up_down_data))
basis.drawing(sess.run(left_right_data))
basis.drawing(sess.run(transposed))
basis_code.basis_encode(up_down_data,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(left_right_data,
basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(transposed,
basis_code.get_encode_path(filepath, 3))
def adjust_saturation(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#将图像的饱和度-5
adjust_less = tf.image.adjust_saturation(img_data, -5)
# 将图像的饱和度+5
adjust_add = tf.image.adjust_saturation(img_data, 5)
#在[-9, 3)的范围内随机调整图像的饱和度
random_adjust = tf.image.random_saturation(img_data, 1, 4)
with tf.Session() as sess:
basis.drawing(sess.run(adjust_less))
basis.drawing(sess.run(adjust_add))
basis.drawing(sess.run(random_adjust))
basis_code.basis_encode(adjust_less,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(adjust_add,
basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(random_adjust,
basis_code.get_encode_path(filepath, 3))
def adjust_contrast(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#将图像的对比度减少0.5倍
adjust_less = tf.image.adjust_contrast(img_data, 0.5)
# 将图像的对比度增加5倍
adjust_add = tf.image.adjust_contrast(img_data, 5)
#在[0.3, 1)的范围内随机调整图像的对比度
random_adjust = tf.image.random_contrast(img_data, 0.3, 1)
with tf.Session() as sess:
basis.drawing(sess.run(adjust_less))
basis.drawing(sess.run(adjust_add))
basis.drawing(sess.run(random_adjust))
basis_code.basis_encode(adjust_less,
basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(adjust_add,
basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(random_adjust,
basis_code.get_encode_path(filepath, 3))
def slice_box(filepath):
img_data = basis_code.basis_decode(filepath)
if img_data.dtype != tf.float32:
img_data = tf.image.convert_image_dtype(img_data, tf.float32)
#调整大小
img_data = tf.image.resize_images(img_data, [180, 267], method=1)
#定义标注框大小
boxes = tf.constant([[[0.13, 0.24, 0.55, 0.89], [0.33, 0.43, 0.48, 0.67]]])
#可以通过提供标注框的方式来告诉随机截取图像的算法那些部分是“有信息量”的.min_object_covered=0.4表示截取部分至少包含某个标注框的40%内容
#bbox_for_draw为重新随机返回的一个标注框大小
begin, size, bbox_for_draw = tf.image.sample_distorted_bounding_box(tf.shape(img_data), bounding_boxes=boxes, min_object_covered=0.4)
#增加一个维度
batch = tf.expand_dims(img_data, 0)
box_data = tf.image.draw_bounding_boxes(batch, bbox_for_draw)
#随机截取出来的图像。因为算法带有随机成分，所以每次得到的结果会有所不同
distorted_image = tf.slice(img_data, begin, size)
with tf.Session() as sess:
basis.drawing(sess.run(img_data))
basis.drawing(sess.run(box_data[0]))
basis.drawing(sess.run(distorted_image))
basis_code.basis_encode(img_data, basis_code.get_encode_path(filepath, 1))
basis_code.basis_encode(box_data[0], basis_code.get_encode_path(filepath, 2))
basis_code.basis_encode(distorted_image, basis_code.get_encode_path(filepath, 3))
image = tf.image.convert_image_dtype(image, tf.float32)
#随机截取图像，减少需要关注的物体大小对图像识别算法的影响
bbox_begin, bbox_size, _ = tf.image.sample_distorted_bounding_box(
tf.shape(image), bounding_boxes=bbox)
distorted_image = tf.slice(image, bbox_begin, bbox_size)
#将随机截取的图像调整为神经网络输入层的大小。大小调整算法是随机选择的
distorted_image = tf.image.resize_images(distorted_image, [height, width],
method=np.random.randint(4))
#随机上下翻转图像
distorted_image = tf.image.random_flip_up_down(distorted_image)
#随机左右翻转图像
distorted_image = tf.image.random_flip_left_right(distorted_image)
#随机使用一种顺序调整图片色彩
distorted_image = distort_color(distorted_image,
color_ordering=np.random.randint(2))
return distorted_image
#通过上面程序，就可以通过一张训练图像衍生出很多训练样本。通过将训练图像预处理，训练得到的神经网络模型可以识别不同大小、方位、色彩等方面的实体
if __name__ == "__main__":
filepath = basis_code.get_andclean_image()
image = basis_code.basis_decode(filepath)
boxes = tf.constant([[[0.13, 0.24, 0.55, 0.89], [0.33, 0.43, 0.48, 0.67]]])
with tf.Session() as sess:
#运行6次获得6种不同的图像
for i in range(6):
result = preprocess_for_train(image, 180, 267, boxes)
basis.drawing(sess.run(result))
basis_code.basis_encode(result,
basis_code.get_encode_path(filepath, i))